The LLM is sampled to make just one-token continuation with the context. Provided a sequence of tokens, only one token is drawn through the distribution of possible following tokens. This token is appended towards the context, and the method is then repeated. Here’s a pseudocode illustr